Prediction Research Of Wheat Aphid Occurrence Dynamic Based On Neural Network And Support Vector Machine

Agricultural pests seriously restrict the yield and quality of crops,forecasting pest accurately is the premise of scientific prevention and control of pests.So it is very important to establish forecast model of accurate and stable in order to control plant diseases and insect pests in advance and reduce crop economic loss.Insect pests forecasting was an application technology which was based on subjects such as biology,ecology and physiology,It forecasted occurrence and degree,distribution trend and so on.As important part of pests management,insect pests forecasting not only can effectively control the pests occurrence development,but also provided good foundation conditions for the agricultural production management and decision.Since the 1950 s,our country suffered several pests outbreak which caused incalculable loss for food production,increasing farmers' income,agricultural development.All along,the agriculture department spend a lot of manpower and material resources and financial resources on the research of insect pest forecasting,in order to reduce loss of farmers,improve crop quality and yield,promote the sustainable development of agriculture.During the past decades,worldwide entomologists have committed to develop various methods for insect pest prediction,including the prediction systems based on personal experience,experiment data or statistic analysis results.Due to the fact of the complex features of insect pest outbreaks,such as unevenness,otherness,diversity,abruptness,and periodicity,traditional linear prediction methods are not able to deliver desirable prediction results.Therefore,modern non-linear theories were employed into pest prediction.Coupled with traditional Kinetic Theory,mathematical statistics and modem computer technology,novel pest prediction methods have been developed,including Artificial Neural Network,phase-space reconstruction,Wavelet Analysis and Support Vector Machine.In this study,it established BPNN?WNN?LSSVM model for forecasting the wheat aphids.1.Forecast model of the maximum rate of wheat aphids atrains based on BPNNIn this study,it based on the maximum rate of insect strains and ten days average temperature?ten days average highest temperature?ten days average lowest temperature?ten days average relative humidity?ten days average precipitation?ten days average sunshine duration?ten days average mean wind speed?temperature and rain coefficient?rain coefficient?temperature and humidity coefficient.It established BPNN model for forecasting the maximum rate of insect strains of wheat aphids during 2007—2011 based on the data of 1980 to 2006.Screened important parameters of BPNN model using trial and error method and made sure 100-13-1 as the structure of the model,the transfer function of hidden layer was logsig and the transfer function of output layer was purelin,the training function was trainlm,weights of network learning function was learngdm,the training number was 3,learning rate was 0.1,expected target minimum error value 10-5,used L-M method to study.Results showed that comparing the prediction results of BPNN and stepwise regression method,the accuracy with BPNN was more than 94.14%and it is better than stepwise regression method which was 75.74%not only the stability but also the accuracy of the model.Therefore,as a result of the BP neural network forecasting model's ability to deal with nonlinear problem,good self-learning,self-organization and adaptive,good generalization ability,which can be used as crop diseases and insect pests forecasting effective method to continue exploring.2.Forecasting model for the occurrence degree of wheat aphids based on PCA-WNNWNN model has the characteristics which include not only analysis ability of wavelet time-frequency but also self-learning,self-organization,nonlinear mapping ability of BPNN.This study was based on the occurrence degree of wheat aphids and meteorological factors from 1980-2014.The study get 9 new affecting factors from 40 basic ones by using of the principal component analysis as the independent variable input model and screened the transfer function and the number of hidden layer nodes by using of the trial and error method.At the same time,the study respectively took the Sigmoid function and Morlet wavelet function as kernel function of BPNN and WNN.Finally,the occurrence degree of wheat aphids forecasting based on PCA-WNN and PCA-BPNN WNN and BPNN model have been established and compared.The study trained the data collected from 1980 to 2009 and forecasted the occurrence degree of wheat aphids from 2009-2014.As the results showed,the four kinds of models can well describe the degree of wheat aphid occurrence;In prediction accuracy and stability,two WNN model are better than the two BPNN models.The PCA-WNN model compared with the WNN model,PCA-BPNN model compared with BPNN model,which was used the principal component analysis to the independent variable dimension in establishing model had precision.As the conclusion showed,PCA-WNN model with the characteristics of time and frequency synchronization analysis and wavelet analysis with neural network self-learning,self-organization,nonlinear mapping ability.PCA-WNN is better than PCA-BPNN not only on the stability but also on the accuracy of the model.The PCA-WNN model can be further explored as a new forecasting method for pests and diseases.3.Forecasting model for the occurrence period of wheat aphids based on SLR-WNNIn this study,it based on the occurrence period and ten days average temperature,ten days average highest temperature?ten days average lowest temperature?ten days average relative humidity?ten days average precipitation?ten days average sunshine duration?ten days average mean wind speed.Selected five meteorological factors from the 70 basic meteorological factor as the independent variable input model by using stepwise regression method.By Screening and made sure the node numbers were 5-15-1,the learning rate 1 was 0.3,learning rate 2 was 0.1.Established the three layers SLR-WNN model,trained the data from 1980 to 2009,forecasted the date from 2009 to 2014.As the results showed,that average prediction accuracy of SLR-WNN model was 93.81%,the average prediction precision of SLR-BPNN model was 91.40%,the average prediction precision of SLR-MLR model was 75.70%.Overall,the prediction accuracy of SLR-WNN model was better than SLR-BPNN model and SLR-MLR model.At the same time,the comparison the MSE of three models,the MSE values of SLR-WNN model was less than SLR-BPNN model and SLR-MLR model,it showed that its stability was better than the other two models.4.Forecasting model for the occurrence amount of wheat aphids based on SLR-LSSVMIn this study,it based on the occurrence amount and five days average temperature?five days average highest temperature?five days average lowest temperature?five days average relative humidity?five days average precipitation?five days average sunshine duration?five days average mean wind speed.Selected 13 meteorological factors from the 70 basic meteorological factor as the independent variable input model by using stepwise regression method.It screened out 10 as the gam,0.01 as the sig2 as the optimal parameter combination by using 5 fold cross-validation method and set up LSSVM model,compared with the BPNN and WNN.Trained data from 1980 to 2009,forecasted the maximum number from 2009 to 2014.As the results showed,the LSSVM model not only can describe the change of the wheat aphid maximum number more precisely,and its average accuracy was 90.03%in five years,it had good prediction accuracy and stability.In previous study,the BPNN and WNN had be proved good prediction effect.But in this chapter,the effect of the two methods were worse than LSSVM in accuracy and stability.Because it had large gap in this study,the largest number of pests was up to 4654 and the lowest number was just 15.6 in 1980-2014 and they had nearly 300 times.It beyond the control of neural network algorithm and lead to serious error.The LSSVM model not only had good generalization ability in treating small sample,high dimension,nonlinear problem like the standard SVM,but also to simplify quadratic programming problem to solve system of linear equations and reduced the computing complexity greatly.It showed that the model had strong generalization ability,good prediction ability,high computational efficiency.It can be used as a new kind of crops diseases and insect pests forecasting method.